Γ-W1-T2: the limiter toggle is a mute, not a crossfade — the ceiling holds across both transitions

The equal-gain dry/wet blend let a peak through at (1-m) of its level. Now the
fade rides only the limited path and the hard edge lands on silence.

src/core/instrument/CLAUDE.md

@@ -292,7 +292,7 @@ anything for a trigger shape.
 - `time_stretch` — the TIME half beside `pitch_shift`'s PITCH half, header-only: `StretchCursor`, the per-output-frame source-feed schedule (a fractional cursor carrying its rate debt, loop-wrapped), plus the rate bounds and their clamp. Rate 1.0 is exactly one source frame per output frame with no residue, which is what makes the unity Preserve read bit-identical to the pre-stretch engine. The bounds are **measured**, not arbitrary — see the header.
 - `velocity_curve` — THE monotone spline, shared by every consumer: the three velocity transfer curves and the three spline EGs. `VelocityCurve` is evaluated as ONE OR MORE Fritsch–Carlson monotone cubic Hermite splines joined at its HARD points — a hard knot is a sub-curve boundary for tangent purposes (exactly what the point array's own ends already are), so the two adjacent segments meet at their natural angle instead of a shared derivative and the no-overshoot guarantee holds PER SEGMENT rather than globally. Points are smooth by default; the ceiling is `kMaxCurvePoints` = 128, a MUSICAL bound (long rhythmic phrases, ~two points per articulation event) and not a performance one — **do not lower it**. `eval(velocity)` is the COLD reader, called once per note-on or once per drawn pixel column; `SplineCursor` is the RT one, an indexed segment search plus one Hermite evaluation with the segment and its tangents cached across samples. Both share the same `segmentTangents`/`hermiteAt` free functions, so there is one spline and not two. It carries its own y `CurveDomain`: UNIPOLAR [0,1] is the amp's GAIN, defaulting to `flat()` (y=1, every velocity→unity — a deliberate non-back-compat replacement of the old fixed `velocity/127` path, Daniel-approved); BIPOLAR [−1,1] is the signed modulation shape for pitch and filter, defaulting to `zero()` so velocity modulates neither until a curve is drawn. A bipolar curve does not imply the absence of a depth beside it: the filter keeps its `velAmount` knob and the two compose multiplicatively (`velAmount × curve.eval(v)`, `play_params.h`), while the pitch curve's throw is the fixed `kVelocityPitchRangeSemitones`.
 - `master_gain` — pure dB↔linear taper math (FB1): normalized [0,1] ↔ dB ↔ linear for the post-mixer master gain control (−∞…+24 dB, norm 0 = true silence, unity ≈ 0.714). Shared by the editor knob and the processor multiply so the needle, persisted value, and audio multiply cannot drift.
-- `limiter` — the master bus's lookahead brickwall limiter, the stage after `master_gain`'s multiply: a 4x-oversampled TRUE-PEAK detector in the SIDECHAIN ONLY (the signal path is never oversampled), one stereo-linked gain, a baked −0.3 dBTP ceiling and **no makeup gain of any kind**. The gain law is a sliding MINIMUM of the per-sample target over the lookahead window followed by a MOVING AVERAGE of the same width: every term of that average is a minimum whose own window contains the sample being gained, so the ceiling is held **structurally** rather than by a tuned attack, and the one-pole release only ever slows the RISE so that bound survives it. Bypassed and settled, `process()` returns without reading or writing a sample — the byte-identical at-rest path, on the same discipline as `live == nullptr` and the filter's exact skip at `modAmount == 0`. `prepare()` owns every allocation and every transcendental; the engage/disengage crossfade is the codebase's standing ramp-every-gain-path-change rule applied to a limiter switching in.
+- `limiter` — the master bus's lookahead brickwall limiter, the stage after `master_gain`'s multiply: a 4x-oversampled TRUE-PEAK detector in the SIDECHAIN ONLY (the signal path is never oversampled), one stereo-linked gain, a baked −0.3 dBTP ceiling and **no makeup gain of any kind**. The gain law is a sliding MINIMUM of the per-sample target over the lookahead window followed by a MOVING AVERAGE of the same width: every term of that average is a minimum whose own window contains the sample being gained, so the ceiling is held **structurally** rather than by a tuned attack, and the one-pole release only ever slows the RISE so that bound survives it. Bypassed and settled, `process()` returns without reading or writing a sample — the byte-identical at-rest path, on the same discipline as `live == nullptr` and the filter's exact skip at `modAmount == 0`. `prepare()` owns every allocation and every transcendental. **Switching is a MUTE, never a blend:** unlimited signal is emitted at weight 1 (the untouched bypass buffer) or at weight 0 and never in between, because a fraction of an unlimited signal is a peak over the ceiling — so the fade always rides the limited path and the hard edge always lands on the bypassed side, against silence. Do not reintroduce an equal-gain dry/wet crossfade over the toggle.
 - `meter_ballistics` — the output meter's UI-side ballistics and dB scale: instantaneous rise, 20 dB/s fall, the 1.5 s peak hold and its release at the same rate, the clip latch, and the dB → normalized map over −60…+6 dBFS. The audio thread publishes raw block peaks and converts nothing; this module is what turns them into what the bar draws.
 
 ### `map/`